A Weakly Supervised Deep Learning Semantic Segmentation Framework

摘要

In this work, we present a weakly supervised deep learning semantic segmentation framework for small-scale image dataset without ground-truth information of segmentation. The main process of this framework is as follows: 1, Training dataset by the region-based convolution neural networks. 2, Getting object classifier and the initial object location. According to the initial location result, 'GrabCut segmentation algorithm is used to iterate the sub-image for object segmentation boundary optimization. The proposed algorithm deals with the image of citrus growth environment and realizes the precise position of citrus in the precise segmentation of citrus object boundary. Extensive experiments show that GrabCut optimal segmentation framework by the region-based convolution neural networks can be used to complete the automatic positioning and segmentation of specific object. The accuracy of our framework achieves 95.8% on the citrus test dataset. Now the framework has been applied to the real citrus grow-detection with stable operation and high accuracy.